Research On Deep Learning Based Biomedical Entity Relation Extraction Algorithm

In the recent years,with the development of biomedical technology,a large number of research results have been produced in biomedical field,and the literature in this field has shown a sharp growth trend.The citations of the medical literature search engine PubMed have exceeded over 30 million and there is rich biomedical knowledge in the massive biomedical literature.Faced with such a large amount of literature,it is increasingly difficult for researchers to obtain useful information through manual reading and understand the latest research progress in biomedical field.Automatic mining knowledge from massive literature through text mining is a feasible way to solve this dilemma,and it has attracted more and more researchers' attention.Entity relation extraction technology is a key technology in text mining field.There are many named entities in the biomedical literature and these entities include proteins,drugs,diseases and et al.There are many relationships between these named entities,such as protein-protein interaction relations,and therapeutic relations between disease and drug.Extracting these entity relations from the biomedical literature by entity relation extraction technology is of great significance to the biologists' systematic biology researches.Currently,there are mainly three categories of methods in this technology: the co-occurrence method,the pattern matching-based method,and the machine learning-based method.Compared with the low precision rate of the co-occurrence method and the low recall rate of the pattern matching-based method,the machine learning-based method has attracted researchers' wide attention due to its excellent performance.The machine learning-based relation extraction methods can be divided into: feature-engineering-based method,kernel-based method,and deep learning-based method.The feature engineering-based method and kernelbased method are dependent on feature design.Because deep learning-based methods do not rely on feature engineering,deep learning-based relation extraction method is the recent research hotspot.Due to the small size of biomedical corpora,most of the current deep learningbased relation extraction methods are based on other natural language processing tools,of which the dependency parsing tree is the most important supplementary feature.This will make the performance of relation extraction model dependent on the performance of these tools.In addition,these tools also limit the prediction speed of the models.Recurrent neural network has been favored by recent work due to its strong sequence model ability.However,because paralleling recurrent neural network is difficult,the inference speed of it is slower than convolutional neural network.In addition,the training of recurrent neural network is also troublesome.Therefore,based on these two problems of existing work,the paper proposes a residual convolutional neural networkbased biomedical entity relation extraction model.The model is implemented based on convolutional neural network and does not rely on any other natural language processing tools.Compared with directly stacking several convolution modules,the residual structure can promote deep neural network's gradient propagation,so that the convolutional neural network can own a deeper architecture.The model was evaluated on several biomedical relation extraction corpora.Compared with the previous models,the model in the paper has achieved satisfactory results.